How ChatGPT, Gemini, Perplexity, Claude and Google AI build an answer

The two ingredients

Parametric knowledge is what a model absorbed during training. It is broad, fluent and frozen at a point in time. It is where a model gets a general sense of what a brand is, what category it belongs to and what it is known for. It is slow to change and hard to influence directly.

Retrieved knowledge is what the system fetches at question time: web pages, news, product data. It is fresher and traceable to specific sources. Citations that appear beside an answer usually come from this step.

Every engine blends the two, and the blend differs by product, by question and sometimes by whether the user switched search on. A question about a long-established brand's history may be answered mostly from the model. A question about this month's pricing needs retrieval.

Google AI Overviews and AI Mode

These are built on Google Search. AI Overviews appear at the top of some results pages; AI Mode is a separate conversational experience. Google has described both as using a technique called query fan-out, in which the system issues multiple related searches across subtopics and data sources and combines what it finds into one response, with links to supporting pages.

Google's documentation says pages must be indexed and eligible to be shown with a snippet to be used as supporting links, and that no additional markup is required. In practice this means good Search fundamentals are the entry ticket, and the sub-questions the system asks decide which pages it reads.

A practical note for measurement: an AI Overview does not appear for every query, and AI Mode is conversational, so a longer natural-language question behaves differently from a short keyword. Test prompts in the form buyers actually use.

Related reading: ChatGPT vs Perplexity citations: how the two differ and Query fan-out: write for the sub-questions an AI answer is built from.

ChatGPT

ChatGPT can answer from the model alone or use search to retrieve current information and show sources. OpenAI operates separate crawlers with separate purposes: one used for collecting data that may be used in training, one that supports search features in ChatGPT, and one used when a user's request triggers a page visit. Site owners can allow or block each in robots.txt.

For brand questions, the retrieved sources are often review sites, publications, directories and the brand's own pages. When search is not used, the answer reflects model knowledge, which can be out of date or incomplete for newer or smaller brands.

Gemini

Gemini, Google's assistant, can ground answers in Google Search results. Because it draws on Google's index, it shares some of the sources that Google's AI features use, but the answers are produced by a different product with its own behaviour.

Google offers a robots.txt control called Google-Extended that lets site owners decide whether content crawled by Google may be used for Gemini model training and grounding. It does not affect whether a page is included in Search.

Perplexity

Perplexity is built around live retrieval. It searches the web for most questions and presents an answer with numbered source citations next to the text. That makes it the most transparent engine to analyse, since the sources are always visible and a brand's chance of being named is tied closely to being present in those retrieved pages.

Perplexity documents its own crawlers, including one for building its search index and one that fetches pages when a user asks a question.

Claude

Claude can answer from model knowledge and can use web search when it is enabled. Anthropic documents separate crawlers for training data collection, for search and for user-initiated fetches, each controllable through robots.txt.

As with the others, the closer a question is to current facts such as pricing, availability and recent comparisons, the more the answer depends on what can be retrieved and how clearly your pages state it.

What this means for your brand

Measure every engine your audience uses, separately. A single blended score hides the fact that you may be strong where retrieval dominates and weak where the model answers from memory, or the reverse.

Fix retrieval first. Make sure your key pages can be crawled by the relevant bots, state facts plainly and keep them current. This helps every engine that searches.

Work on the model's general sense of you over the longer term: consistent facts across the web, credible third-party coverage, and clear descriptions of what you are and who you are for.

OClever tracks AI assistants including ChatGPT, Gemini and Perplexity side by side and stores each full answer with its citations, so differences like these show up in your own data.